Page 2: Introduction and an overview of machine learning
Today, the games industry is one of the fastest growing and most adaptable industries, utilizing cutting edge technology in order to push the boundaries of gaming. With the global video gaming market exceeding that of television, film and music (1). This being said, a wide multitude of technology is beginning to enter widespread use in games design – Machine learning is just one of these. In this paper, I will cover the past, current and the predicted future uses for machine learning in the games and film industry. In conjunction, I will discuss, explain, and evaluate machine learning in games design.
Machine learning is a system which utilizes artificial intelligence to improve, adapt and learn from collected data without being pre-programmed to behave in such a way (2). To be put simply, machine learning allows a system to change behaviour, learning in a comparable manner to a human.
The learning which takes place is initiated by data given to the system, or data which the system has collected itself. To which then, the system uses the data to make a decision or determination. This output is often based on a correct example output which will be provided to the system as a reference (2).
While being a technically specific label when referring to the application of artificial intelligence, machine learning is a rather broad term which may involve a wide range of different methods and subsystems.
These methods include, but are not limited to:
Supervised learning algorithms: Systems which use supervised learning utilize an observation, much like a rule the machine should adhere to, provided by a human to link an input to an output. For example, a system would use the “round and yellow” which the system will be told to interpret as a “tennis ball” using the predetermination already supplied by the human (2).
Unsupervised learning algorithms: The systems which utilize unsupervised learning rely on data input which had been given to the system or data which has been collected by the system but without the pre-attached connections thus having the machine to make comparisons and linking collected data. An example of could be simply illustrated with a machine that has a cache of street-view images. Which the machine then makes a connection between similar features shared by the images which it will continue to recognize from that point (2).
Reinforcement Learning algorithms: Reinforcement learning, being more complex than both supervised and unsupervised learning is in actual fact more greatly comparable to that of a more human like learning through trial and error. Reinforcement learning involves two variables the agent and the reward. The first being the failure condition and the second which can be best described as a quantitative reward for the system. The system then tries hundreds of different approaches to reach the predetermined reward. The amount of steps involved to get to the reward are deducted from it at the end. The system then takes the number of steps and deducts them from the reward to which then the machine choses the most effective route. This way the machine learns in a manner which is similar to that of trial and error (3).
Page 3: History of machine learning
It is widely recognized that the foundations for machine learning had been built upon early neural network models dating back to the late 1950’s with the first artificial neural network, ‘Perception’ being invented which utilized basic pattern recognition (4).
It wouldn’t be until the late nineties where machine learning would creep its way back into public interest when the IBM chess playing computer (5), Deep Blue had beaten Garry Kasparov in a standard chess match (6, 7).
Going into the 21st century, it would be another decade until any noticeably big leap or milestone achieved by machine learning would come to light.
Which was when in 2012 GoogleBrain was created, utilizing complex learning algorithms and Google’s databases to categorize images based on common visual traits (2). Making GoogleBrain stand out as the deep learning systems had access to amounts of data which AI had previously hadn’t been able to touch (5). GoogleBrain acted as a catalyst in a time when there was a great deal of uncertainty in regards to the practical applications of machine learning. Allowing other high profile corporations to invest and develop their own rivaling and advanced machine learning systems.
The next big corporate debut of machine learning was unveiled by Facebook in 2014 with their release and implementation of DeepFace. A machine learning algorithm which uses photos from FaceBook’s data libraries to recognize individual people in photographs (2). With FaceBook AI research scientists claiming “Our method reaches an accuracy of 97.35%” (8) indicating an accuracy higher than or equal to that of a human.
Moving into 2015, an explosion of private ventures into machine learning had taken place. Headlining with Open AI from Elon Musk. As well as Amazon’s own machine learning platform most noticeably found in its Alexa product and Microsoft’s machine learning documentation being released (5).
Between 2015 and now (The first quarter of 2019) the AI revolution has been been spearheaded by machine learning beginning to influence more humanitarian and social contexts. With AI translators having higher accuracy than human translators in some cases. Domestic smart products and features such as Alexa, Google Assistant, Cortona, Siri (9). And most unexpectedly, machine learning had begun to infiltrate some of the most creative of scenes, such as the film industry and the video gaming industry.
Page 4: Contemporary uses for machine learning in games design
The games industry is the largest media industry, being larger than that of movie, TV and music (1). This makes the industry a competitive market where businesses must utilize cutting edge technology in order to retain a monopoly and reduce production costs. A technology which has blown up in the recent decade is that of machine learning. To which the games industry is fully embracing and using as standard.
Currently, machine learning is used in a wide variety of ways in the industry to reduce expenditure, man hours and improve quality. These applications include but are not limited to;
Developing artificial intelligence, level design, texturing and PBR mapping, idea concepting and idea visualization, render quality improvement, simulating human trial and error designing.
All of the above listed examples are currently used in the games industry and are available commercially or a specific to certain studios. All of these jobs aren’t unique to machine learning and can be carried out by human artists. However, an artificial intelligence can do the same jobs in a fraction of the time and requires very little to no expenditure (10).
The following is a list of examples and how machine learning is currently influencing them.
Automatically developing artificial intelligence: With machine learning, specifically utilizing reinforcement learning algorithms as mentioned previously in the paper. Artificial intelligence can discover the path of least resistance through attempting a multitude of different approaches and then evaluate which is the most effective; fulfilling the success criteria while also avoiding any failure conditions and being the least process intensive. This way, an artificial intelligence can evolve and adapt with minimal input from a coding perspective, thus reducing time and money required to make a complex behavior system (1).
This being said, there are still many unforeseen errors a machine learning system could encounter and anomalous conclusions which may have to be rectified by a human.
AI driven level design: Artificial intelligence and machine learning in level design have played a big role in the industry, debuting with some of the earliest examples of randomly generated game worlds. This often works by the system being fed examples of desired or real world images of terrain, to which the machine will be told to adhere to iconic characteristics of these terrains, and then generate a terrain of its own based on the data it had been given (10). This is used for games such as Space Engineers.
Of course, having a computer generate terrain presents its own issues and problems. Some of these may include an inability to recreate the random characteristics of nature, as well as being unable to distinguish the difference between subtle and prominent details.
Machine learning PBR mapping: With softwares emerging such as Substance Painter, Substance Designer and Z-Brush. Mesh painting is becoming standardized within the games industry. But more interestingly, these softwares include functionalities to calculate where wear, rust, moisture, snow or any other kind of surface disruption is likely to occur. It can do so by drawing from a library of real world examples which it then uses to project a similar effect onto 3D meshes (10). There is little drawback to systems such as this as it is very simple for artists to make amendments to PBR maps.
Page 5: Future uses for machine learning in games design
In the near future, it is often speculated that machine learning will exist within every stage of game development, and subsequently, every other artistic medium. To which I am in agreement.
It is of my own personal projection that in the future, machine learning will be utilized to replace much of the lower skilled jobs in the games industry. Jobs such as recreating real world objects, humanoid and animal like animations, idea generation and basic problem solving. This then leaves artists with a much more authoritative job, comparable to that of a producer or lead artist. Having to govern and approve of the outputs of machine learning.
Page 5 continued: Proposed issues with machine learning
It is often believed that machine learning systems will replace jobs. Potentially steal the fun from artistic work. Or simply be less efficient than a human artist doing the same job
To which I can empathize and understand. But ultimately conclude that these matters of opposition are of no concern for the following reasons.
Firstly, I believe that given assistance from AI will create new jobs in developing and working with such machine learning systems. As well as alleviating the overall workload placed on any studio which utilizes machine learning assisted development.
Secondly I believe that machine learning assisted development will remove the tedious elements from the creative process and instead allow artists to spend more time being creative and innovating (10).
Page 6: Personal research undertaken
[Figure 1]
For this paper, I had conducted some of my own research by creating a survey which was to gain statistical data for preferred games development approaches and artistic methods. Above is an example of the data which I had gathered from the results I had gotten from the survey. This data suggests that there is still balanced uncertainty in regards to the involvement of artificial intelligence in conjunction to human artistic development.
My data however isn’t conclusive as the study group was limited both numerically and in diversity and so the results that were drawn from the survey aren’t conclusive to a broad enough extent as would be necessary in determining a wider opinion.
Bibliography
John Stephenson, 29th November, 2018, 6 Ways Machine Learning Will Be Used In Games Development, logikk.com https://www.logikk.com/articles/machine-learning-in-game-development/
Hope Reese – Brandon Vigliarolo, 9th April, 2018, Machine Learning: A Cheat Sheet, TechRepublic.com
https://www.techrepublic.com/article/machine-learning-the-smart-persons-guide/
Prateek Bajaj, Reinforcement Learning, GeeksforGeeks.org
https://www.geeksforgeeks.org/what-is-reinforcement-learning/
Daniel Mayer, 2018, What Is Machine Learning? A Definition, ExpertSystem.com
https://www.expertsystem.com/machine-learning-definition/
Hugo Mayo – Hashan Punchihewa – Julie Emile – Jack Morrison, 2018, History Of Machine Learning, AIinRadiology.co.uk
https://www.doc.ic.ac.uk/~jce317/history-machine-learning.html
BBC, 2019, The History Of Machine Learning, BBC.com
https://www.bbc.com/timelines/zypd97h
Wikipedia authors, last edited in 2019, Garry Kasparov, Wikipedia.org
https://en.wikipedia.org/wiki/Garry_Kasparov
Yaniv Taigman – Ming Yang – Marc’Aurelio Ranzato – Lior Wolf, 14th June 2014, DeepFace: Closing the Gap to Human-Level Performance in Face Verification, research.fb.com
https://research.fb.com/publications/deepface-closing-the-gap-to-human-level-performance-in-face-verification/
Iliya Valchanov, 25th January 2018, Machine Learning: An Overview, datascience.com
https://www.datascience.com/blog/machine-learning-overview
Andrew Price, 25th October 2018, The Next Leap: How A.I. will change the 3D industry, Blender, youtube.com
https://www.youtube.com/watch?v=FlgLxSLsYWQ
My Games Design Blog 